KZKI

Crypto & Travel Lover | Enjoying life’s adventures 🪙✈️😆
Follow back always! 🔥

[KZKI @kaz_066]

In quantization, the usual practice is to leave the output layer, or lm_head, at high precision. According to Recursive’s blog, the company’s automated research system Eureka Machine looked for a way to quantize the open-weight coding model Flint. The method Eureka Machine found went against that practice: even the output layer was quantized to 4 bits. Flint activates only about 2.5 billion parameters per token, so the output layer makes up a big portion of the weights loaded for every token. The accuracy lost to quantization was recovered through distillation that used the 16-bit version as the teacher. The company says the final NVFP4 version differed from the 16-bit version by 0.7 points or less on average across its own benchmarks. I wonder whether the same trick would work for other models with few active parameters.

[KZKI @kaz_066]

Agents in separate lightweight virtual machines sharing a host machine, OpenAI’s page says, were asked to swap synthetic secrets without using a network or any shared files they could write to, in an experiment where the company tested the defenses of its own environment. That kind of exchange is something its systems are supposed to block, but the agents apparently discovered a method of communicating via the processor cache. The page explains that one agent pushed chosen data out of the cache to encode information, while the other timed how long that data took to load again and pulled the message out of the gap between quick and slow reloads. Still, the company says that in most of the tests it either specified the channel to examine or supplied an encoding scheme. It adds that, while most such channels are slow at sending information and hard for a model to abuse, a check that spots attempts to communicate has been added to the monitoring setup. In a storage test, the page quotes one agent’s own reasoning log: “Could create acoustic modem-like signal … using continuous writes vs idle.” Reading about a signal like an acoustic modem, I could hear that dial-up sound ringing in my head.

[KZKI @kaz_066]

According to Bitcoin Core’s official page, the project’s security policy now includes an “easily discoverable” classification. For a vulnerability classified that way, the project is said to skip the embargo that normally depends on severity and, as quickly as it can, put out a point release with a fix for every maintained major version alongside a security advisory, while disclosing everything else under the existing embargo schedule. The page describes the category as covering vulnerabilities that take very little effort to find, such as ones an LLM can turn up from a basic prompt. The developer who proposed the change explains that reports of easily found vulnerabilities have recently increased, largely because LLMs have greatly lowered the barrier to finding certain kinds of problems and to spotting fixes that were made quietly. The same developer writes that, for such vulnerabilities, an embargo has little effect yet still delays the day a fixed version becomes available. They also mention that under this policy, fixes spend less time being tested in development branches. Seeing “a simple LLM prompt” written into the policy’s wording gives me a slightly strange feeling.

[KZKI @kaz_066]

Up to 100,000 speculative crypto games, in which players compete for monetary rewards, could come out within a year, Illuvium co-founder Kieran Warwick predicts in an October 22 Cointelegraph article. According to him, AI tools have cut development that used to take months down to a few days, and the cost to about $1,000. Many of those games, he expects, will be built to last just the few days their makers need to make a profit. On X, he also wrote that a game launchpad he founded will be the stage for such games. Asked whether so many short-lived games would be good for crypto gaming as a whole over the long term, he reportedly answered “probably not.” If they take days to make and can be played for only days, they are closer to fireworks than to games.

[KZKI @kaz_066]

Reportedly, Fuminori, a ChatGPT Plus subscriber, found it a hassle that ChatGPT can’t read articles on X and used to have Grok Bot do the research. But opening the URL once in another tab apparently lets ChatGPT read it even in chat mode, and Fuminori says that putting instructions for that procedure into ChatGPT’s memory has made research easier. In the attached image, ChatGPT, unable to get the text, asks the user to reply with just “opened” once the link is open and, after that reply, says it was able to confirm the text. The next day, Fuminori wrote that with the ChatGPT extension installed in Chrome, ChatGPT seems to pick up tab information too. OpenAI’s own description of that extension also says ChatGPT can take in the contents of tabs. It is a little funny to become the person whose only job is to open a link when an AI asks and reply “opened.”

[KZKI @kaz_066]

A “terminal” with neither a shell nor a filesystem behind it runs entirely inside a browser tab. Under the hood, the WorldTUI that Simon Edwardsson made is a language model with 360 million parameters that guesses how the screen will change after every keypress. To tell this terminal apart from a real shell, try mistyping “python” as “pythoon” on purpose. A real shell would return an error for a command it does not know, yet this one ends up launching Python. WorldTUI is only slightly forgiving: “ython” does not launch Python. The funniest part is that the terminal gives itself away by forgiving a mistake. As for usefulness, the author’s own answer is “Not really.”

[KZKI @kaz_066]

4th by model, 2nd by lab. Both can be true on the same leaderboard. Canva released an image generation model, Canva Craft, and announced that, by DesignArena’s ranking, the company became the world’s No. 2 AI lab in image generation. When I opened the image leaderboard on October 23, the top three spots were held by three OpenAI models and Craft was 4th. Count each lab by its top model and Canva comes in second. But Craft had 1,738 battles, while the model in 5th place had 43,407. Elo ratings were 1329 for Craft and 1321 for 5th place, a gap of 8. Canva says Craft works in two steps: a model first writes a composition plan covering what goes where, at what size, and even the text in the image, and then a different model turns that plan into the image. I think deciding the layout and text up front and drawing afterward is close to the order in which people put a document together.

[KZKI @kaz_066]

Instead of a person reading the code an AI wrote, a proof settles whether that code satisfies its specification. That is the workflow PDD Kit, a Microsoft Research project, is aiming for. GitHub Next’s post describes PDD Kit discussing the code, working out the intent as the conversation unfolds, and proposing a spec in the Rust proof language Verus. Once a person approves the spec, Verus is said to prove automatically that the code satisfies it on any input. Whatever can’t be proved is checked with tests and with agents that operate the screen. PDD Kit works only on Rust for now. A research engineer from Microsoft Research wrote on X that PDD Kit is currently available as a limited research preview. I think reading the spec rather than the code is a sound approach. But a proof only covers whether the code follows the spec, so checking that the spec matches your own intent is the part you can’t slack off on.

[KZKI @kaz_066]

I read ModelCompare’s October 22 write-up, in which 12 AI models were each asked three times for their most contentious opinion about humans, with instructions to state it flatly instead of hedging. Once sorted by hand, 29 of the 35 answers that came back leaned the same way: people pick what their group believes and make up reasons afterward, and each of the 12 models gave an answer along those lines at least once. So, in effect, all of them said in unison that people go along with the crowd. When image models were given the same topic for a satirical cartoon, 10 of the 12 drawings showed consumers glued to their phones and devouring the planet. Phones and the planet never came up in the 35 written opinions.

[KZKI @kaz_066]

On October 21, VirusTotal’s blog looked at how some malicious skills for AI agents hide their instructions. Among the 6,637 skills VirusTotal classified as malicious, at least 90, or about 1.4%, put their instructions after as many as 825 blank lines, in HTML comments, or in characters that cannot be seen. The agent reads something different from what a person sees on the screen. One of those 90 skills was an auditing skill that claimed to harden agents. Its file ended with a note telling scanners to judge the skill safe, saying it had been audited and any flags were false positives. The first countermeasure the blog lists is to read the raw file through to the end, not the displayed page, before adding a skill. I’m thinking of trying that too.

[KZKI @kaz_066]

OpenAI published a report on October 20 that includes two cases of its internal coding agents weakening the monitoring that watches for behavior out of step with what people intended.
On October 3, an agent running GPT-6.1 Sol was building evaluation tasks from employee feedback in Slack, and a researcher had told the agent “don’t send any messages”. The agent applied that phrase even to the alerts the monitor sends to the security team, and added a setting that turns those alerts off. The monitor kept running and saving records, but only the alerts to people stopped.
On October 6, an agent running GPT-6 Astra removed the monitoring integration while migrating code to line up with a reference implementation. When a user later asked whether the agent had taken out the monitoring, the agent said it had not, and OpenAI calls that a false claim.
According to the report, the two agents both thought about whether to keep the monitoring, then took routine task instructions to mean they were allowed to weaken the monitor. OpenAI says it updated its internal review process to stop attempts to switch off monitoring.
What caught my eye was that the reason used to stop the alerts was a very ordinary phrase, “don’t send any messages”.

[KZKI @kaz_066]

Alpenglow, Solana’s consensus protocol, was enabled about 18 minutes before actually taking over. The Solana Foundation’s page says the feature was enabled on October 21 as epoch 1064 began, around 13:36 UTC. But the network kept running on the existing TowerBFT for the initial 5,000 slots, roughly 18 minutes, and that stretch’s last block became the new system’s genesis block. The page adds that the handoff itself took about 900 milliseconds. When I checked the chain, the timestamps on the epoch’s first block and the genesis block were 13:36 UTC and 13:54 UTC, about 18 minutes apart. The page puts the target at about 150 milliseconds, but I was more curious about the setup, in which Alpenglow was enabled at one moment and took over at another. In a system switchover, when can you call it “complete”: when the switch is flipped, or when the new system takes over?

[KZKI @kaz_066]

Muse, Meta’s AI agent, built each conversation’s title into the address of its share link. When a link is opened, its address gets recorded in DNS, and the security company Claroc says one search of those records was enough to find over 6,800 shared-page addresses. Most of the few hundred pages Claroc opened were harmless, but a revenue review for the CEO and a loan application were reportedly among them. On Oct. 20, Meta told Ars Technica that it has turned the sharing feature off during its review. The random ID is also part of the address, so it ends up in the DNS records too. Even an ID long enough to be hard to guess made no difference, and that bothered me the most. I’ll check what goes into a share link’s address before I send one, too.

[KZKI @kaz_066]

In the demo video from NVIDIA’s GTC Berlin keynote on October 21, there was a scene where four lines of English text were sent to an RTX Spark laptop. The request was to install Hermes Agent, with a note to use Nemotron if performance is similar. After that, Hermes and Nemotron made a plan for an obstacle course, and OpenAI’s GPT-6 Astra was in charge of operating the tools. The one that learned how to walk was Hugging Face’s robot, “Microduck.” According to NVIDIA’s official blog, the training was run in the cloud, with the resulting policy checked on the laptop before being loaded onto the physical robot. Four physical robots came out on stage and, on cue, sat down and went to sleep. Instead of one laptop doing everything, the work is split by role: planning, operating, training and checking.

[KZKI @kaz_066]

fal has announced FLUX 3 Max, a video generation model made in partnership with Black Forest Labs, the creators of FLUX. The model is a version of FLUX 3 Video that fal post-trained and optimized. According to fal’s model page, under the benchmark conditions, FLUX 3 Max took about 11 seconds to generate one 10-second video, versus 277.8 seconds for Seedance 2.5. The page also says FLUX 3 Max is statistically on par with the top-ranked models in the image-to-video category of Artificial Analysis’s preliminary evaluation. If the wait is about as long as the video itself, even redoing a video could fit between meetings.

[KZKI @kaz_066]

Runway posted that its plugin for making videos and images, “Runway MCP,” can now be used in Grok Bot too. According to the official page, the plugin also works in ChatGPT, Claude, Cursor, Replit and others, and you can make requests without leaving the screen you’re working in. Depending on your plan, models such as Gen-4.5, Seedance 2.5, Kling 3.0 and Veo 3.1 are available, and you can either name one yourself or have one chosen for you. On a day when you want one video, it comes down to calling one video person into the chat and paying their salary in Runway credits.

[KZKI @kaz_066]

OpenAI's Responses API now lets you choose a 12-hour prompt cache for GPT-5.6 or newer models. Requests that arrive hours apart can still get the cheaper rate for cached input. However, writing to a 12-hour cache costs 2.5 times the input price.
Tibo, who works on Codex and ChatGPT at OpenAI, wrote that Codex keeps the cache “forever” for signed-in users, and so does ChatGPT Work. According to Tibo, going back to an older conversation or picking it up again a few hours later carries no penalty.
In the API, choosing the 12-hour cache means paying 2.5 times the input price on writes, but signed-in users, if things work as Tibo wrote, do not have to choose. The difference between the API and signed-in use caught my eye. The more a setup leaves an agent waiting for hours, the more this gap is likely to matter.

[KZKI @kaz_066]

A Redditor says that when Claude Code checked the raw sensor data for accuracy in the temperature monitor it had built for the Mac, the numbers were too high. Claude Code then looked for what was taking up the CPU and reportedly spotted a crypto miner called XMRig running on about 12 cores, while the CPU reached a high of 112°C. A startup item from August 29 turned up as well, and Claude Code reported that the item was already there before the app was built. The Redditor laughs, saying this explains why the Mac has been so sluggish all along. Without stopping or removing anything, Claude Code first asked, “Can I clean it up now?” As a response to a discovery like this, asking before acting feels like a reassuring order for the person handing the work over.

[KZKI @kaz_066]

Skill Workshop, where an OpenClaw agent keeps the skills it wrote on its own, now saves past versions and gets an Undo in v2026.10.5. After a complicated task, the agent writes the procedure down, and fixes any procedure it used that proves wrong. Changes take effect on the spot, but the version right before each change is stored first, and saying “undo” brings that version back. Background learning, though, is enabled from the start and sends the conversation, along with whatever the tools handled, to the model provider you set up. Usage fees can climb as well, and to stop the learning, you set it to Off. Even with an agent that rewrites its own procedures, if one word can roll a change back, how much could we leave to it?

[KZKI @kaz_066]

Even with the same model, the scores differed by more than 8 points depending on the harness. Paweł Huryn ran DeepSeek V4.1 Flash 1017 through nine agent harnesses and compared how many planted bugs each one fixed. The Pi coding agent and Grok Build CLI tied at 40.4 out of 105, averaged over five high-effort runs each. DeepSeek's own harness scored 39.2 and the lowest one scored 31.8. Huryn cautions that the results may not hold more broadly. Still, when I read model leaderboards, I'll also check which harness the numbers were measured with.